240 likes | 320 Views
Unsupervised Information Extraction from Unstructured, Ungrammatical Data Sources on the World Wide Web. Mathew Michelson and Craig A. Knoblock. Abstract. Extracting unstructured data is difficult. Traditional methods do not apply. Solution: Unsupervised extraction.
E N D
Unsupervised Information Extraction from Unstructured, Ungrammatical Data Sources on the World Wide Web Mathew Michelson and Craig A. Knoblock
Abstract • Extracting unstructured data is difficult • Traditional methods do not apply • Solution: Unsupervised extraction • Results are competitive to supervised methods
Introduction • Web data could be useful if extracted (i.e Craigslist)
Introduction • Posts are not structured • The Phoebus method works on this data • But it requires much user input (supervised) • The paper presents and optional unsupervised method • This work extends on unsupervised semantic annotation
Introduction • Current work on UIE relies on redundancy • This approach does not use structural assumptions • This approach relies on similarity no redundancy • This approach creates relational data
Unsupervised Extraction Steps of the algorithm: • Automatically choosing the Reference Set • Matching Posts to the Reference Set • Unsupervised Extraction
Unsupervised Extraction • Automatically choosing the Reference Sets - They choose a reference set based on similarity - They calculate a similarity score and sort the sets - They use percent difference and average score - The algorithm scales linearly with size - They use multiple metrics as similarity score
Unsupervised Extraction • Matching Posts to the Reference Set - A vector-space model is used to match posts - The Jaro-Winkler metric is used to match tokens - Attributes that do not agree are removed - Now we can query the posts (Yay !)
Unsupervised Extraction • Unsupervised Extraction - A baseline is created between extracted field and reference set field - We remove tokens based on the baseline
Experimental Results Reference Sets Post Sets
Experimental Results Jensen-Shannon similarity
Experimental Results TF/IDF similarity
Experimental Results Jaccard similarity
Experimental Results Jaro-Winkler TF/IDF similarity
Experimental Results Results
Experimental Results Dice similarity
Experimental Results Jaccard similarity
Experimental Results TF/IDF similarity
Experimental Results Dice vs Phoebus
Experimental Results Jaro-Winkler vs Smith-Waterman
Experimental Results Comparison with other methods
Related Work • SemTag is a similar system • But it uses a crafted taxonomy • In contrast, SemTag focuses on disambiguation • CRAM is also similar but it requires labeling
Conclusion • This paper introduces an unsupervised information extraction technique • The Jensen-Shannon distance metric is better • Using text acronyms would be beneficial • Entity extraction could be a good idea